Multi-agent systems rely on communication to complete tasks, and latent communication passes continuous representations to reduce text generation and re-encoding. However, multi-agent systems must reconfigure agent models, whereas existing latent-communication methods are built for fixed model combinations, so their interfaces may fail after replacement. Our pilot study confirms this interface failure and identifies the requirement for interface reuse after model replacement. We present LatBridge, which connects agent models through a shared Bridge and separate interfaces under a shared latent-communication protocol. After agent model replacement, LatBridge reuses unchanged interfaces, selects matching low-rank interfaces, and combines them through the Bridge, training only missing ones with all models frozen. The Bridge processes source representations in the basic mode and constructs target input without autoregressive generation in the accelerated mode. Experiments support effective communication after model replacement, with mean basic-mode accuracies of 14.8% on MATH and 27.3% on ALFWorld next-action prediction across eight small-model configurations. Analyses examine the factors supporting task accuracy and communication efficiency.
Figures & tables
Figure 1: Agent model replacement in latent communication. Interfaces in latent-communication methods can fail after replacement. LatBridge keeps the Bridge model fixed, reuses the unchanged interface, and selects or trains a matching interface for the new model.
Task
Target
Original
No communication
Replacement
MATH
Qwen
12.00±0.00
0.50±0.00
0.67±0.24
Llama
5.00±0.00
0.00±0.00
0.50±0.41
ALFWorld
Qwen
23.63±2.53
0.25±0.36
0.00±0.00
Llama
19.41±4.09
0.00±0.00
0.00±0.00
Table 1: Pilot study of source-agent model replacement based on Interlat ( Du et al., 2026 ) . Accuracies are percentages reported as mean ± standard deviation.
Figure 2: Workflow of LatBridge . (1) Matching interfaces are retrieved and composed after agent model replacement. (2) Only missing interfaces are trained and stored. (3) Latent communication around the Bridge uses either the basic or accelerated mode.
Bridge
Source
Target
Qwen2.5-0.5B / Llama-3.2-1B
Qwen2.5-7B / Llama-3.1-8B
Basic mode
Accelerated mode
Basic mode
Accelerated mode
MATH
ALFWorld
MATH
ALFWorld
MATH
ALFWorld
MATH
ALFWorld
Qwen
Qwen
Qwen
16.0
33.6
16.5
24.6
30.0
26.9
57.0
45.5
Qwen
Llama
11.0
32.8
16.0
42.5
39.0
24.6
44.0
44.8
Llama
Qwen
14.0
31.3
16.5
24.6
28.5
30.6
57.0
45.5
Llama
Llama
18.0
21.6
16.0
42.5
38.5
27.6
44.0
44.8
Table 2: Task accuracy (%) across eight model configurations at two model scales, grouped by Bridge model.
Figure 3: Task accuracy under Bridge-output replacement in the basic mode . Thin curves show model configurations, labeled by source and Bridge models. Thick curves show target-family means with bootstrap 95% confidence bands.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Bridge
Source
Target
MATH
ALFWorld
Qwen
Qwen
Qwen
1.39
1.81
Qwen
Qwen
Llama
1.07
1.59
Qwen
Llama
Qwen
1.37
1.44
Qwen
Llama
Llama
1.08
1.52
Llama
Qwen
Qwen
1.21
1.33
Llama
Qwen
Llama
1.78
1.59
Appendix
Table 3: Mean teacher-forced target loss for the large-model basic mode . Each value uses the same examples, interfaces, Bridge outputs, and target-side inputs as the corresponding large-model accuracy result in Table 2 .
Replacement case
Input interface
Output interface
New interfaces
Small (K)
Large (K)
Source: no adaptation
Reuse mismatched
Reuse matched
0
0
0
Source: interface available
Select matched
Reuse matched
0
0
0
Target: interface available
Reuse matched
Select matched
0
0
0
Source: interface missing
Train matched
Reuse matched
1
57–131
229–262
Target: interface missing
Reuse matched
Train matched
1
57–131
229–262
Appendix
Table 4: Interface operations after agent model replacement . Parameter counts are in thousands for rank-32 interfaces and exclude the frozen agent and Bridge-model backbones. Stored matching interfaces are reused, and only missing interfaces are trained.
Figure 4: Target-side representation-space analysis on held-out ALFWorld tasks. Columns separate target-agent families, while rows show target-native prompt references, Qwen Bridge outputs, and Llama Bridge outputs. Colors denote task type and marker shapes denote the source-agent model. PCA is fit separately within each target family. Each row uses an independent local scale with coordinate values retained on the axes.
Scale
Task
Before
After
Reduction
Interlat (end-to-end)
–
ALFWorld
9,190
390
23.6×
LatBridge (local Bridge)
Small
MATH
1,262.53
0.334
3,777×
Small
ALFWorld
559.84
0.335
1,669×
Large
MATH
2,256.96
0.367
6,149×
Appendix
Table 5: Mean communication latency (ms). LatBridge compares basic and accelerated modes, averaged over eight configurations. Interlat compares full and 8-step latent generation ( Du et al., 2026 ) . Reductions use each system’s own timing boundary.
Scale
Bridge
Source
Target
MATH
ALFWorld
Basic mode
Accelerated mode
Basic mode
Accelerated mode
Small
Qwen
Qwen
Qwen
1545.78
0.257
651.82
0.257
Small
Qwen
Qwen
Llama
1534.54
0.404
649.37
0.384
Small
Qwen
Llama
Qwen
1508.14
0.268
655.67
0.269
Small
Qwen
Llama
Llama
1499.46
0.401
572.73
0.398
Small
Llama
Qwen
Qwen
987.26
0.404
573.44
0.402
Appendix
Table 6: Per-configuration Bridge processing latency in milliseconds. Each value is the mean over the fixed task test set under the synchronized timing boundary in Appendix B .
Variant
Qwen–Qwen–Llama, MATH
Llama–Qwen–Qwen, ALFWorld
Accuracy (%)
Parameters (K)
Accuracy (%)
Parameters (K)
LatBridge (basic mode)
11.0
151.6
22.4
188.4
Without two-stage training
2.5
151.6
33.6
188.4
Without decoupled output training
3.0
94.2
0.0
94.2
Without input interface
2.5
94.2
0.0
94.2
Without output interface
0.5
57.3
1.5
94.2
Appendix
Table 7: Small-model component and training ablations on two representative configurations. Configurations are listed as Bridge, source, and target models. Learned-parameter counts include all parameters updated by each variant; the final row additionally updates low-rank parameters in the Bridge model. Bold marks the best accuracy for each task.
Figure 5: Task accuracy under different communication budgets in the basic mode . We retain different prefixes of the Bridge output before the output interface. Thin curves show source–Bridge pairs, with circles and squares denoting Qwen and Llama targets. Thick curves show target-family means with bootstrap 95% confidence bands.
Analysis
Qwen
Llama
Cross-source retrieval
76.0
88.5
Basic-mode probe
50.5±3.7
56.5±5.6
Accelerated-mode probe
73.0±7.3
71.5±4.6
Cross-mode retrieval
38.5
44.5
Appendix
Table 8: Representation-analysis accuracy (%) by Bridge model. Probe values are five-fold means ± standard deviations.